A Video-based Fall Detection Network by Spatio-temporal Joint-point Model on Edge Devices

Tripping or falling is among the top threats inelderly healthcare, and the development of automatic fall detectionsystems are of considerable importance. With the fast developmentof the Internet of Things (IOT), camera vision-based solutions havedrawn much attention in recent years. The traditional fall videoanalysis on the cloud has significant communication overhead.This work introduces a fast and lightweight video fall detectionnetwork based on a spatio-temporal joint-point model to overcomethese hurdles. Instead of detecting falling motion by the traditionalConvolutional Neural Networks (CNNs), we propose a LongShort-Term Memory (LSTM) model based ontime-series joint-point features, extracted from apose extractorand then filteredfrom ageometric joint-point filter. Experiments are conducted toverify the proposed framework, which shows a high sensitivityof98.46%on Multiple Cameras Fall Dataset and100%on URFall Dataset. Furthermore, our model can achieve pose estimationtasks simultaneously, attaining73.3mAP in the COCO keypointchallenge dataset, which outperforms the OpenPose work by8%.